Existential types: Unused closure environment is part of concrete type

[zroug]

Compare

#![feature(existential_type)]

existential type T: ;
fn test(_unused: impl Copy) -> T {
    || {}
}

https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=194cca6bdb23ec0b634bc01dea43abd4

which results in error: type parameter `impl Copy` is part of concrete type but not used in parameter list for existential type to

#![feature(existential_type)]

existential type T: ;
fn test(_unused: impl Copy) -> T {
    fn workaround() -> impl Fn() {
        || {}
    }
    workaround()
}

https://play.rust-lang.org/?version=nightly&mode=debug&edition=2018&gist=23789461920ea49d85042a9604f86f97

which compiles fine.

There are quite a few open issues for existential types so I hope I haven't overlooked a duplicate. #42940 might be related.

Meta

rustc 1.33.0-nightly (03acbd71c 2019-01-14)
binary: rustc
commit-hash: 03acbd71c977cd63ce5f39ba9b6fe9ffd578785a
commit-date: 2019-01-14
host: x86_64-unknown-linux-gnu
release: 1.33.0-nightly
LLVM version: 8.0

[Centril]

cc @cramertj @alexreg @nikomatsakis

[alexreg]

I'm confident this is a duplicate of #42940, but let's wait for others to confirm.

[cramertj]

This isn't a duplicate of that issue-- the error message here is "correct" in a sense. The issue here is that the closure being generated is still generic over type parameters, even when they're unused. We don't have any logic today (AFAIK) for detecting when a type parameter goes unused in a closure, so by default they capture all generic parameters in scope. Desugaring the argument-position impl Copy makes this more obvious:

#![feature(existential_type)]

existential type T: ;
fn test<A>() -> T {
    || {}
}

error: type parameter `A` is part of concrete type but not used in parameter list for existential type
 --> src/main.rs:4:19
  |
4 |   fn test<A>() -> T {
  |  ___________________^
5 | |     || {}
6 | | }
  | |_^

The fix here is to detect when a type parameter is unused in a closure and make it so that the closure is no longer generic over that parameter. I remember talking with @nikomatsakis about this in the past and I think they indicated that this was hard to do in the current system.

[nikomatsakis]

That would indeed be quite hard to do. For one thing, the set of generics is determined quite early on -- this would require us to wait until we see what the closure does. But also, it's non-trivial to figure out when the type parameter was important to the closure.

[cramertj]

I guess I'll close this for now as "working as intended", though it'd definitely be nice to drop unnecessary type parameters from closures anyways (for codegen purposes).